DEPARTMENT OF ECONOMICS UNIVERSITY OF CYPRUS

Love thy Neighbor, Love thy Kin: Voting Biases in the Eurovision Song Contest Sofronis Clerides and Thanasis Stengos Discussion Paper 2006-01

P.O. Box 20537, 1678 Nicosia, CYPRUS Tel.: ++357-2-892430, Fax: ++357-2-892432 Web site: http://www.econ.ucy.ac.cy

Love thy Neighbor, Love thy Kin:

Voting Biases in the Eurovision Song Contest∗

Sofronis CleridesThanasis Stengos

February 2006

Abstract

The Eurovision Song Contest provides a setting where Europeans can express their sen-timents about other countries without regard to political sensitivities. Analyzing votingdata from the 25 contests between 1981-2005, we find strong evidence for the existence ofclusters of countries that systematically exchange votes regardless of the quality of theirentries. Cultural, geographic, economic and political factors are important determinants ofpoints awarded from one country to another. Other non-quality related factors such as orderof appearance, the language of the song and the gender of the performing artist, are alsoimportant. There is also a substantial host country effect.

Keywords: Eurovision, social networks, games of trust.

JEL Classification: Z13.

∗The research assistance of Adamos Adamou and Gavriella Panayiotou is gratefully acknowledged. We aresolely responsible for any errors.

Author affiliations and emails: Clerides: University of Cyprus and CEPR; s.clerides@ucy.ac.cy; Stengos:University of Guelph; thanasis@uoguelph.ca. Corresponding author: Sofronis Clerides, Department of Eco-nomics, University of Cyprus, P.O.Box 20537, CY-1678 Nicosia, Cyprus. Phone: +357-22892450; fax: +357-22892432.

1 Introduction

One Saturday in May of every year millions of Europeans sit glued in front of their television

to watch one of the entertainment highlights of the year: the Eurovision Song Contest (ESC).

The contest has evolved from a humble, seven-country music festival that was first staged in

1956 to a glitzy extravaganza that is expected to include 40 participating countries in 2006 and

reach a potential audience of one billion people worldwide. Each country is represented in the

contest with one song and then votes for the best song among all other countries’ entries. The

country-centered format of the ESC has turned it into a national contest where people root for

the country’s entry in the same way that they root for their national football team.

For this reason, but also because musically the contest is mediocre at best, many people would

argue that the most exciting part of the ESC is not the singing but the voting. After all the songs

have been performed live on stage, the presenters get on the phone with a representative of each

participating country and the country’s votes are read out. The procedure is suspenseful and

exciting. Eurovision buffs are well aware of the fact that votes are not always cast on the basis

of a song’s quality. Friendly countries always exchange votes, while no country expects to get

points from a country that it is not on good terms with, no matter the quality of the song. This

setup makes the ESC a fascinating experiment where each country can express its sentiments

about another without having to give regard to the political sensitivities and considerations that

usually accompany inter-governmental relations.

In our paper we use data from all contests between 1981-2005 to examine cultural, geographic,

political and economic factors as possible explanations of voting behavior that results in the

formation of voting blocks and social networks. Large and systematic differences in voting

patterns may reflect, in addition to differences in tastes, some deeper sociological likes and

dislikes among nations. These systematic biases that go well beyond the aesthetic quality of

the song itself are captured by what we have identified as affinity factors, that is variables that

measure how each country feels towards another country. We find that affinity variables are

very important in the voting process. We also test for the presence and importance of other

1

attributes in the song delivery such as the order of appearance, the language of the song and

the gender of the performing artist, as well as whether there is a benefit for the country hosting

the competition. These variables also turn out to be quite important in explaining voting

patterns. The econometric analysis is conducted in two ways. We first analyze the ordered

ranked preferences of each participant and then we proceed to examine the scores that each

participant gives to the other participating countries. Since the score data are censored we use

estimation methods that account for that.

The paper is organized as follows. In the next section we present a brief history of the

ESC. We then proceed to present an overview of the existing literature followed by a description

of the voting patterns that govern the contest. The description of the data, the econometric

methodology used and the results from our analysis for different specifications that we tried are

presented in section five. Finally, in the next section we offer some concluding remarks.

2 The Eurovision Song Contest

The (ESC) originates from a contest that first took place in 1956 under the name Eurovision

Grand Prix which, in turn, was modeled after the popular Italian San Remo Festival. The

first contest took place in Lugano, Switzerland with only seven participants. The number of

participants has grown now to 39 in the 2005 competition on the 50th anniversary of the ESC.

The list of participants includes countries that do not belong geographically in Europe such as

Israel and Morocco, while many other non-European countries are petitioning to enter.1

The contest is run by the European Broadcasting Union, an association that is mostly com-

prised of European national television broadcasters. Although the specific rules have changed

over the years, the basic format of the competition remains unchanged. Each national broad-

caster submits an entry that will represent the country in the contest. The selection process is

up to the broadcaster; the only restriction is it has to be an original song.2 On the night of the1Armenia is scheduled to be the 40th participant in 2006.2For many years there was also a restriction that the song had to be performed in one of the country’s official

languages. This was lifted because of complaints that it put some countries at a disadvantage, a view that is

2

event the competing songs are performed live and then each country awards points to songs by

other countries. The country that earns the most points wins the contest and gets to host the

following year’s event.

Irrespective of whether ESC actually contributes to the creation of better music in the

continent, the festival itself constitutes an example of a truly international forum where a country

can express an opinion about another country, free of political or economic considerations. The

awarding of points to songs goes beyond rewarding a “good” song, since in that case a good

song should receive the same number of votes from all other countries barring any major taste

differences. In addition to differences in tastes, some deeper sociological likes and dislikes among

nations that manifest themselves in systematic biased voting. If it were that all participants

were equally likely to produce good and bad songs then the distribution of votes from a certain

contestant to her fellow competitors should be more or less equal over time. Systematic biases

however, conceal the different considerations beyond the aesthetic quality of the song itself that

enter the voting preferences of participant countries.

The basic structure of the voting system has been in place since 1975, whereby each voting

country awards points to ten other countries, itself not included. There have been various

changes over the years to the main format of the contest, the structure of the voting system, the

number of contestants and the character of the songs. The main changes that are important for

our analysis are as follows:

(i) In 1975 the current scoring system was introduced. Each country ranks the ten best songs.

The top-rated song gets 12 points, the second gets 10 and the next eight get 8, 7, 6, ..., 1.

(ii) An important change was the introduction of televoting. Up until 1997 each country’s

votes were decided by a panel of experts. Televoting was allowed as an option in 1998 and

was soon adopted by all participating countries.

(iii) In order to accommodate more contestants the competition was split into two rounds in

2004. Countries with poor records in the contest have to compete in a qualifying round

supported by the results in this paper.

3

which determines the songs that will compete in the main event. The countries involved

in the qualifiers can vote in the final, but only those that reach the final can receive votes.

This allowed the total number of participants to increase from 24 in 2003 to 36 in 2004

and 39 in 2005.

3 Related literature

A network is a mechanism of exchanges between participants, whereby these exchanges may

represent information in the form of communication messages between workers (Gandal, King,

and Van Alstyne 2005) or scientists collaborating on a joint project (Newman 2003). In a typical

social network individuals are depicted as nodes and the links between the nodes represent

communication exchanges between these individuals. Over time new nodes and new links will

appear as new individuals enter the network and new collaborations are established. In the

context of the ESC a network can be established as exchange takes place between different

countries in the form of points. The countries are the individual nodes and the connecting

links between these nodes represent the points exchanged between the countries. These edges or

connecting links can be directed, undirected, weighted and unweighted, depending on whether

these links are taking into account the direction of the vote from country A to country B or not,

as well as whether the number of points exchanged is accounted for in the depicted link or not.

The minimum degree of connectivity (number of connecting links) that a node (country) can

have in a given year of competition is ten, as each country assigns points to ten other countries.

The maximum degree is equal to the total number of the other contestants in that year because

the country can receive points from all other countries.

Fenn, Suleman, Efstathiou, and Johnson (2005) use complex dynamic networks to analyze

the voting patterns of the ESC in the period 1992 to 2003. The authors are able to uncover

nonlinear patterns that emerge over time that contradict the hypothesis that the ESC is a

random contest. In a random contest countries simply vote for the best song without any likes

or dislikes for the other contestants. If a country is equally capable of producing a good or

4

bad song as any other country that would generate a pattern over time that would not differ

from a random number generator simulating the results of the competition over time. Fenn et

al (2005) find this is not the case and there are patterns that are not compatible with random

behavior. In social networks two nodes that are connected to a third one are more likely to be

linked together as well, as someone’s acquaintances are more likely to know and communicate

with each other. In the context of ESC, the same phenomenon may occur in the form of “voting

blocks” within the contest where there are clustering effects that differ from the pattern that

would arise in a “random contest” environment.

Doosje and Alexander (2005) examined the issue of reciprocity in voting behavior between

countries in the ESC and concluded that countries give on average more points to countries from

who they tend to receive higher scores. In other words reciprocity between participant countries

over time acts as a catalyst for vote clustering. Bornhorst, Ichino, Karl, and Winter (2004) use

experimental data to look at the impact of cultural diversity on agents’ choices of partners as

well as on the outcomes of economic interactions. In a dynamic trust game environment, where

subjects were divided between cultural lines, northern and southern Europeans, they show the

existence of cultural biases in the way agents conduct their economic activities. In the context of

the game northerners seem to be culturally biased against southerners, as they perceive them as

less trustworthy, where trust is measured by the tendency to reciprocate by making a generous

payback for a transfer received. The results are not due to stereotyping, since they emerge and

are reinforced by repeated interactions between different nationalities even when agents are not

characterized by strong stereotyping at the outset of the interaction. Other examples of trust

games in the economic literature are Fershtman and Gneezy (2001) who examined the issue

of trust between Ashkenazi (Jews of European descent) and Sephardi (Jews of Middle Eastern

origin) Israelis, in the context of an one shot game. In general similar conclusions are reached by

Guiso, Sapienza, and Zingales (2003) who find that trust differences between people of different

countries affect the level of trade.

In the context of the ESC voting biases are the equivalent of transfer biases in the above

games. Cultural (trust) biases between countries (groups) would produce the same clustering

5

effects that are predicted by a non-random contest environment. In this paper, we use the work

of Fenn et al (2005) on the formation of voting networks over time as a starting point. Our

intention is to offer statistical evidence as to the factors that help create these networks, by

identifying the variables that help explain the voting preferences of the participant countries.

Ginsburgh and Noury (2004) and Haan, Dijkstra, and Dijkstra (2005) also examined different

factors that affected voting behavior in the ESC. Ginsburgh and Noury (2004) looked at the

possible effect of vote trading or logrolling as opposed to an index of quality of their song,

whereas Haan, Dijkstra, and Dijkstra (2005) looked at the possible differences of the judging

behavior of expert juries as compared to that of the general public using the effect of the order

of appearance on the outcome of the vote in data sets from the finals of the ESC as well as

those from national finals. In the ESC until 1998 voting was conducted by expert juries whereas

in national competitions the public was typically more involved in the song selection. The

findings of Ginsburgh and Noury (2004) find some evidence that language affects voting but

other measures of culture are either statistically insignificant or very close to zero numerically.

Haan, Dijkstra, and Dijkstra (2005) find evidence that order of appearance has an effect on voting

although this effect is smaller for expert juries. In our paper using a data set that includes all

voting records between participants in the ESC for the period 1981-2005 we examine the link

between cultural, geographical and economic factors as possible explanations of the behavior

that results in the formation of possible voting blocks and social networks.

4 Voting patterns

We have collected voting data from all Eurovision song contests from 1981 to 2005.3 Each

contest featured 18-26 countries; a total of 39 countries have participated in the ESC in the

period under examination.4 There are a total of 656 country pairs that are observed between 1

and 25 times; only Spain, Sweden and the UK participated in all 25 contests.3The data are readily available at the official ESC website (http://www.eurovision.tv) as well as various

websites maintained by ESC aficionados (e.g. http://www.kolumbus.fi/jarpen/ and http://www.esctoday.

com).4For the years 2004 and 2005 we use the countries that participated in the final round only.

6

The first thing to do is to look at the mean number of points awarded between different

countries and look for country pairs that exhibit “abnormal” voting behavior. In order to set

a benchmark suppose that song quality is randomly determined every year. Each country gets

a draw from the same distribution so that expected song quality is the same for all countries.

Each country gives out a total of 58 (= 12 + 10 + 8 + 7 + 6 + 5 + 4 + 3 + 2 + 1) points and also

expects to receive, on average 58 points. In a contest with N participants, the expected number

of points received per participant is 58/(N − 1). Given that the median contest size is 23 and

assuming random song quality, we expect that each country will receive from each other roughly

58/22 = 2.64 points.

In order to compare this with our data we calculated the mean number of points awarded

between the members of each pair in our sample. In Table 1 we present the pairs with the highest

and lowest averages; we excluded pairs that appear fewer than three times. Column (4) reports

the mean number of points awarded from country A to country B while column (5) reports

the mean number of points awarded in the opposite direction. We see that many countries

systematically give to specific other countries many more points than the 2.64 we expect on

average. At the same time, those countries also receive a lot of points from the countries they

give to. The correlation between columns (4) and (5) is .8757. Give and thou shalt receive?

Perhaps.

In reality, of course, song quality is not random. Some countries have stronger musical

traditions or more mature entertainment industries and are able to consistently produce above

average songs. In columns (6) and (7) we present the mean number of points received by each

country from all other countries. To ensure comparability we took this mean over contests that

included both countries. Looking at the first entry in the table, in the 18 contests that both

Cyprus and Greece competed, each country averaged 2.2 points. Yet Cyprus gave Greece an

average of 10.9 points and received an average of 10.1. The difference between 10.9 and 2.2 is

the “overgiving” from Cyprus to Greece after controlling (crudely) for song quality. Cyprus has

given to Greece an average of 8.7 points more than the average Greece has received from all

other countries. Greece and Cyprus are the most extreme example, but it is clear from the table

7

Table 1: Pairs with the highest and lowest point exchangesMean points awarded “Overgiving”

Country A Country B Obs. A to B B to A all to A all to B A to B B to A(1) (2) (3) (4) (5) (6) (7) (8) (9)

I love you:Cyprus Greece 18 10.9 10.1 2.2 2.2 7.9 8.7

Slovakia Malta 3 11.3 6.7 0.3 4.4 6.4 6.9Ukraine Russia 3 8.7 7.3 2.9 2.6 4.4 6.1Estonia Latvia 4 9.3 8.5 4.5 3.4 4.0 5.9Finland Estonia 5 8.4 4.4 0.6 2.7 3.8 5.7Croatia FYRMacedonia 5 6.6 8.0 2.8 1.0 5.2 5.6Turkey Bosnia 10 6.5 4.6 2.8 1.4 1.8 5.1Croatia Bosnia 11 6.5 6.1 2.3 1.4 3.8 5.1

Denmark Sweden 20 8.2 6.1 3.0 3.8 3.1 4.4Slovenia Croatia 9 7.1 5.0 1.5 2.8 3.5 4.3Romania FYRMacedonia 5 5.0 5.2 1.7 1.1 3.5 3.9Germany Poland 8 5.8 4.3 2.9 1.9 1.4 3.9

Estonia Sweden 9 7.0 5.9 3.4 3.6 2.5 3.4Norway Sweden 24 7.1 4.7 2.5 3.7 2.2 3.4Norway Denmark 19 6.6 3.9 2.3 3.2 1.6 3.4Cyprus Serbia 12 6.6 4.7 2.6 3.4 2.1 3.2

Romania Russia 6 5.3 5.0 1.9 3.3 3.1 2.0Croatia Malta 13 6.3 4.8 2.6 3.3 2.2 3.0Sweden Iceland 17 4.9 5.5 3.7 1.9 1.8 3.0

Romania Greece 6 6.0 3.8 1.8 3.2 2.0 2.8

I love you not:Romania Latvia 4 0.0 0.0 2.9 4.4 -2.9 -4.4Croatia Sweden 13 0.5 0.5 2.8 3.8 -2.3 -3.3Turkey Latvia 5 1.2 0.0 3.1 3.6 -3.1 -2.4

Denmark Croatia 8 0.0 1.5 3.7 2.8 -2.2 -2.8Malta France 15 0.2 2.0 3.7 2.9 -1.7 -2.7

Correlations: .8757 -.0328 .9527

8

that overgiving occurs between other countries also. In the bottom part of the table we see pairs

of countries that do not give many points to each other. There are sone notable differences here

also although they are not as large because 2.64 is much closer to zero than it is to 12. Note that

controlling for quality has actually increased the correlation to a striking .9527. Going beyond

simply looking at country pairs, just reading out loud the names of the countries that appear

in Table 1 gives some clues of more interesting behavior. Countries in Scandinavia, the Balkans

and eastern Europe dominate the table. Croatia appears six times, Sweden five, Romania four,

Estonia and Latvia three. Lots of points are exchanged within three clusters: nordic countries;

balkan countries; and countries hailing from the former Soviet Union.

A helpful way of displaying voting alliances is through the use of network graphs. In Figure

1 we display such a graph.5 Countries were arranged in approximate geographical position with

such adjustments made to make the patterns clearer and keep the graph size small. A connecting

link from country A to country B was drawn if country A gave to country B at least 6.1 points

(the mean plus two standard deviations) on average over all contests they both participated in.

From the figure we can discern certain network clusters that mimic the geographic positions of

the countries involved. These separate clusters include the nordic group of countries (Sweden,

Finland, Norway, Denmark, Iceland and Estonia); the former Soviet republics (Estonia, Latvia,

Lithuania, Moldova, Russia, Ukraine); and the former Yugoslav republics (Bosnia-Herzegovina,

Croatia, FYR Macedonia, Slovenia, Serbia & Montenegro).

5 Econometric analysis

5.1 Conceptual framework

The basic problem faced by each country is to identify and rank the ten best songs in a contest.

Our basic modeling assumption is that this decision will depend on two factors: affinity that

each country feels towards each other country and the perceived quality of each song. The latter5The graph was created using the NetDraw program which is part of the UCINET package. We are grateful

to Neil Gandal for pointing us to this literature.

9

Figure 1: Network graph depicting overgiving

can be further decomposed into objective quality that relates to observable song attributes and

subjective quality that relates to a country’s idiosyncratic preferences for a certain type of song.

In order to be concrete, let aij denote the affinity between country i and country j. The

quality of country j’s entry as perceived by country i is denoted by qij = θj+εij , where θj denotes

objective song attributes and εij denotes country i’s idiosyncratic preference for country j’s song.

The overall valuation of each song amounts to a mapping from these three factors to an one-

dimensional index vij = f(aij , θj , εij). The songs are then ranked: the song j with the highest

vij is ranked 1st (RANKij = 1), the next one 2nd (RANKij = 2), and so on up to the 10th

song. The remaining songs are not ranked. RANKij is then translated into points POINTSij

as described above. Thus variable POINTSij takes values in {0, 1, 2, 3, 4, 5, 6, 7, 8, 10, 12} and

variable RANKij takes values in {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, .}, where a dot “.” signifies entries that

10

are not ranked. RANKij is the natural dependent variable to use but is also cumbersome and

difficult to interpret. For that reason we will rely mostly on specifications that use POINTSij

as the dependent variable.

We parameterize affinity and objective quality as aij =∑

l βl xijl and θj =∑

m γm wjm

respectively. Assuming a linear specification for f() yields the following expression for a song’s

valuation:

vij =L∑

l=1

βl xijl +M∑

m=1

γm wjm + εij . (1)

5.2 Explanatory variables

Our goal is to determine the factors the determine affinity aij and perceived quality qij and

the relative importance of each in determining the outcome of the voting process. A useful

point of reference is the gravity model that has been used extensively in the international trade

literature to model trade flows between countries.6 In a typical specification the variable to

be explained is the amount of exports from country A to country B. Explanatory variables

include size, geographic proximity and variables that aim to capture cultural, religious and

political links between the two countries; in short, variables that capture affinity. In our case

the dependent variable will be the number of points given by country A to country B (or the

rank assigned to it). Our choice of explanatory variables was guided by the gravity equation

literature. One notable difference is that we use trade flows (the dependent variable in the

gravity equation) as an independent variable to capture the closeness of economic relations

between the two countries. We defined two trade variables; EXPORTS for the pair (A,B)

is the percentage of country A’s exports that went to country B in 1994; TOTTRADE is

the sum of the values of the EXPORTS variables for the two countries in each pair; that is,

TOTTRADEij = TOTTRADEji = EXPORTSij + EXPORTSji. TOTTRADE captures

overall importance of trade between the two countries while EXPORTS captures asymmetric6Anderson (1979) provides a theoretical justification for the gravity model; Leamer and Levinsohn (1995)

review the empirical literature.

11

effects that may arise from exports as opposed to imports.7

In order to capture geographic closeness we created the variable PROXIMITY, defined as the

negative of the distance between airports in capital cities unless the countries have a common

border in which case PROXIMITY was set to zero. In that case we capture the pure neigh-

borhood effect that is not contaminated by a case where the capitals are far apart in distance,

yet the countries have common borders.8 In order to capture the effect of a common language

we created two dummy variables. MAINLANG takes the value of 1 if the main language of the

two countries is the same (e.g. Germany and Austria, Greece and Cyprus). ANYLANG takes

the value of 1 if the two countries have any one common language, even if it is not the main

one (e.g. Italy and Switzerland get a 0 for MAINLANG and a 1 for ANYLANG). The variable

ANYRELIG was defined in the same way as ANYLANG to capture similarities in religious

beliefs.

We also collected data on several observable characteristics that may be thought to impact

a song’s perceived quality. One thing we can not measure is pure artistic quality.9 What we can

do is test various conjectures that have been circulating in Eurovision circles.10 For example, it

is widely thought that the language that the song is written in is important; specifically, songs in

“strange” languages are at a disadvantage. We construct two dummy variables, ENGLISH and

FRENCH to identify songs written in those two languages. Female performers are said to do

better than males; the dummy variables SOLOMALE and SOLOFEMALE aim to capture that

effect. We also include a variable named DUET to capture differential effects of dues versus

larger groups. Order of appearance is also rumored to be important; the variables FIRST3

and LAST3 identify the first three and last three songs in order of appearance. Finally, the

hosting country typically does better than average, so we include the dummy HOST which7We selected an indicative year in order to reduce the data collection burden. It should be a sufficiently good

proxy for our purposes. Trade data came from the IMF’s Direction of Trade Statistics database.8Distances were obtained from the distance calculator at http://www.etn.nl. Country pairs with common

borders, language and religion were identified from information given in The CIA Factbook, http://www.cia.gov/cia/publications/factbook.

9A cynic would say that such a measure would take the value of zero for all entries and thus would not beidentified.

10The webmaster of http://www.kolumbus.fi/jarpen/ lists some of these conjectures and provides supportingevidence using data from winning entries only.

12

Table 2: Descriptive statisticsVariable Mean Std dev. Min. Max.POINTS 2.677 3.675 0 12EXPORTS 0.028 0.051 0 0.389TOTTRADE 0.056 0.077 0 0.448ANYRELIG 0.358 0.479 0 1ANYLANG 0.222 0.415 0 1PROXIMITY -4.634 1.509 -6.268 0HOST 0.042 0.2 0 1ENGLISH 0.271 0.444 0 1FRENCH 0.12 0.325 0 1FIRST3 0.132 0.338 0 1LAST3 0.132 0.338 0 1SOLOMALE 0.267 0.442 0 1SOLOFEMALE 0.441 0.496 0 1DUET 0.126 0.332 0 1NORDIC 0.033 0.18 0 1SOVIET 0.005 0.073 0 1YUGOSLAV 0.007 0.084 0 1PASTGIVING 0.515 2.091 -7.33 11.9PASTGIVSQ 4.639 11.55 0 142.1

takes the value one if the receiving country hosts the competition. These are variables that can

be construed as packaging effects of the song delivery. The first three are variables where the

performing country has a choice in affecting the delivery of its song by choosing the artist and

the language of delivery, whereas the last three variables capture the exogenous characteristics

of the song delivery which are beyond the performer’s control. Table 2 provides descriptive

statistics for all the variables used in the analysis.

5.3 Estimation and results

Our data are essentially an unbalanced panel of partial rankings by each participant of all other

participants. Such data are not common in economics and we will draw on techniques developed

in sociology in order to analyze them.11 An observation in our data is uniquely defined by the

triple giver-receiver-year. We note that every country pair appears twice in each year. For11See Allison and Christakis (1994) and Marden (1995).

13

example, the pair (A,B) appears once when A gives to B and once when B gives to A. Thus in

a contest with Nt participants we have Nt · (Nt − 1) observations. A total of 656 country pairs

appear in all; the total number of observations we work with is 12,151.

The rank-ordered logit specification, also known in the literature as exploded logit and

as the Plackett-Luce model (Marden 1995), is the appropriate method to use in the context

of a contest with many participants, where some of these participants will be ranked. The

method was proposed in the econometrics literature by Beggs, Cardell, and Hausman (1981)

and further refined Hausman and Ruud (1987) who also coined the name rank-ordered logit.

Note that there is more than one ranked participant as voting does not simply pick the best

among the contestants but also ranks the rest in the group of alternatives. This is different

than the standard ordered logit or probit specifications, where one chooses one among many

alternatives. The rank-ordered logit has a sequential interpretation. One first chooses the best

among all participants (assigns the maximum vote of 12). Next, the voting country selects the

best alternative among the remaining alternatives and the process continues in that way. The

decisions at each stage of the decision making process are described by a standard ordered probit

or logit model.12 Note that the model also accounts for censoring as only the top ten alternatives

are ranked among more than twice as many contestants. The method of estimation is maximum

likelihood (Beggs, Cardell, and Hausman 1981).13

A unique aspect of our data is that we have repeated rankings by the same contestants. The

entries that each contestant has to rank differ from year to year but also have common elements

(the affinity aspect). We are not aware of any previous work that uses the rank-ordered logit

in a similar setting. In order to keep the likelihood function tractable we make the simplifying

assumption that there is no link between rating countries in different years. In other words,

France in 1985 is a different country than France in 1986. This means that we can not fully12Ginsburgh and Noury (2004) used a ordered probit model as a way of checking the robustness of their results,

whereas Haan, Dijkstra, and Dijkstra (2005) used linear OLS methods to analyze the rankings.13The implementation of the maximum likelihood estimator is obtained as the partial likelihood estimator of

an appropriate Cox regression model for waiting time (Allison and Christakis 1994). In this case, higher valuesof an alternative is equivalent to a higher hazard rate of failure. In other words a higher stated preference isrepresented by a shorter waiting time until failure.

14

exploit the panel aspect of our data. The basic model that we estimate is

RANKijt = f(xij,wjt; θ) (2)

where vector xij includes variables that capture affinity between countries i and j, wjt includes

characteristics of entry j and θ is the parameter vector to be estimated. Note that xij has no t

subscript as all our affinity variables are time-invariant.

The rank-ordered logit has the disadvantage that the results do not have a natural inter-

pretation. For this reason we also estimated an alternative specification using POINTSit as

the dependent variable. We treat POINTSit as a continuous variable and apply the Tobit

correction to account for censoring. The basic specification is:14

POINTSijt = α + x′ijβ + w′jtγ + εijt (3)

In Table 3 we present the results of our base specification using four methods: the rank-

ordered logit, the simple tobit, the fixed-effects panel tobit and the random effects panel tobit. In

all cases the set of affinity variables are very important in explaining voting behavior. EXPORTS,

TOTTRADE MAINLANG, and MAINRELIG are statistically significant and positively affect

the votes that one country gives to another. This is of course in line with the argument that

cultural affinity plays a very strong role in affecting exchange behavior expressed in the formation

of networks as argued by Fenn et al (2005). PROXIMITY also has a positive impact that is

statistically significant but less so than in the case of the other variables. This suggests that

geographic proximity is less important than cultural factors. The affinity variables do not appear

important only in the case of the fixed effects specification (except for EXPORTS ) as the latter

are another way of capturing the time-invariant characteristics embodied in the affinity variables.

From Table 3 it is also evident that the factors that affect the presentation of the song14Ginsburgh and Noury (2004) used the observation of the pair indexed by the giving country as that of the

dependent variable, but included the observation where the same country is a receiver as a regressor to capturepossible voting exchange or reciprocity among the members of the pair. That introduces endogeneity in theestimation that needs to be accounted for by the use of instrumental variables.

15

Table 3: Results from the base specificationVariables Rank- Panel tobit

ordered Pooled random fixedlogit tobit effects effects

EXPORTS -0.270 6.519∗∗ 5.159∗ 6.013∗∗

(-0.664) (-1.975) (-2.241) (-1.843)

TOTTRADE 1.211∗ 2.389† 5.274∗∗ 11.3(-0.592) (-1.45) (-1.614) (-21.2)

ANYRELIG 0.177∗∗ 0.840∗∗ 0.611∗∗ -1.527(-0.038) (-0.156) (-0.169) (-4.202)

ANYLANG 0.065 0.563∗∗ 0.570∗∗ 1.496†

(-0.040) (-0.175) (-0.192) (-0.768)

PROXIMITY 0.034∗ 0.139∗ 0.141∗ 1.745†

(-0.014) (-0.058) (-0.063) (-0.968)

HOST 0.534∗∗ 3.185∗∗ 2.615∗∗ 2.540∗∗

(-0.061) (-0.319) (-0.298) (-0.302)

ENGLISH 0.542∗∗ 1.576∗∗ 1.054∗∗ 1.453∗∗

(-0.047) (-0.157) (-0.155) (-0.172)

FRENCH 0.020 0.230 0.844∗∗ 0.633∗

(-0.046) (-0.217) (-0.231) (-0.259)

FIRST3 -0.160∗∗ -0.803∗∗ -0.783∗∗ -0.900∗∗

(-0.042) (-0.205) (-0.191) (-0.193)

LAST3 0.192∗∗ 0.956∗∗ 0.890∗∗ 0.997∗∗

(-0.039) (-0.199) (-0.185) (-0.189)

SOLOMALE -0.147∗∗ -0.622∗∗ -0.584∗∗ -0.521∗

(-0.045) (-0.215) (-0.201) (-0.206)

SOLOFEMALE 0.206∗∗ 1.036∗∗ 1.124∗∗ 1.221∗∗

(-0.042) (-0.198) (-0.185) (-0.189)

DUET 0.237∗∗ 1.504∗∗ 1.294∗∗ 1.374∗∗

(-0.051) (-0.251) (-0.235) (-0.24)

INTERCEPT -1.494∗∗ 0.382 7.26(-0.379) (-0.402) (-6.511)

Obs. 12,151 12,151 12,151 12,151ln L -15,475 -22,844 -23,192 -21,971χ2 520.24 606.81 525.62 2352.57ρ 0.145(s.e.) (0.008)

Standard errors in parentheses.Significance levels: † : 10%, ∗ : 5%, ∗∗ : 1%.

16

(packaging effects) are also very important. If the language of performance is ENGLISH, the

song has a definite advantage over a performance in another language. What we found striking

is the fact that the order of appearance is very important in how a song is judged. Participants

that perform last (the last three positions in the order of appearance) do significantly better

on average than those who perform first, even though the order of appearance is determined

randomly, confirming the findings of Haan, Dijkstra, and Dijkstra (2005). One possible argument

for this result is that presence of a “fading memory effect”, where jurors seem to better remember

the performance of the later contestants and fail to recall as well those who have performed

earlier. The HOST country receives about three extra points from each participant. We think

that at least two factors contribute to this large impact. One is that the host is rewarded for

going through the trouble of “putting up the show”. The second is that the host country invests

more heavily in the contest because it wants to make a strong showing on its home turf.15 A

similar host country effect was obtained by Bernard and Busse (2004) in their study of medals

won in Olympic Games. Finally, the presence of female artists gives an advantage to the song

of the country in question by more than one point on average, indicating a certain bias of jurors

or of the voting public towards female performers. Duets also seem to be better appreciated by

voters than larger groups.

The results are qualitatively very similar across specifications. In the remainder of the

analysis we present results using the panel tobit model which is more flexible and affords a

natural interpretation of estimated coefficients as points.

One possible criticism of our base specifications is that the variables measuring trade volumes

are endogenous. Indeed, the whole point of the gravity model is to explain trade flows using

our other independent variables. Including trade flows in the model may be soaking up some

of the impact of other factors. To test for this we re-estimated the model without the trade

variables. We report the results in column 2 of Table 4, alongside the results from the base

specification which we reproduce in column 1 to facilitate comparisons. Removing the trade

variables naturally leads to a lower maximal point on the likelihood function while it also in-15Greece, the host of the 2006 ESC, has just announced that it has recruited Anna Vissi, the country’s most

successful performer over the last 20 years, to represent the country in this year’s contest.

17

Table 4: Results from alternative specificationsVariables (1) (2) (3) (4) (5)EXPORTS 5.159∗ 5.055∗ 4.01† 3.957†

(-2.241) (2.207) (2.29) (2.259)

TOTTRADE 5.274∗∗ 6.393∗∗ 6.97∗∗ 7.818∗∗

(-1.614) (1.599) (1.673) (1.66)

RELIGION 0.611∗∗ 0.795∗∗ 0.511∗∗ 0.433∗ 0.291(-0.169) (0.167) (0.173) (0.178) (0.182)

LANGUAGE 0.570∗∗ 0.599∗∗ 0.241 0.618∗∗ 0.302(-0.192) (0.191) (0.192) (0.198) (0.200)

PROXIMITY 0.141∗ 0.333∗∗ -0.002 0.024 -0.09(-0.063) (0.056) (0.065) (0.068) (0.069)

HOST 2.615∗∗ 2.622∗∗ 2.618∗∗ 2.498∗∗ 2.492∗∗

(-0.298) (0.298) (0.298) (0.310) (0.310)

ENGLISH 1.054∗∗ 1.071∗∗ 1.067∗∗ 0.875∗∗ 0.896∗∗

(-0.155) (0.155) (0.155) (0.162) (0.162)

FRENCH 0.844∗∗ 0.924∗∗ 1.058∗∗ 0.686∗∗ 0.903∗∗

(-0.231) (0.230) (0.230) (0.240) (0.24)

FIRST3 -0.783∗∗ -0.761∗∗ -0.747∗∗ -0.877∗∗ -0.846∗∗

(-0.191) (0.191) (0.191) (0.201) (0.201)

LAST3 0.890∗∗ 0.885∗∗ 0.886∗∗ 0.675∗∗ 0.671∗∗

(-0.185) (0.185) (0.185) (0.199) (0.199)

SOLOMALE -0.584∗∗ -0.596∗∗ -0.579∗∗ -0.468∗ -0.449∗

(-0.201) (0.201) (0.201) (0.219) (0.219)

SOLOFEMALE 1.124∗∗ 1.119∗∗ 1.128∗∗ 1.114∗∗ 1.122∗∗

(-0.185) (0.185) (0.185) (0.201) (0.201)

DUET 1.294∗∗ 1.293∗∗ 1.259∗∗ 1.297∗∗ 1.257∗∗

(-0.235) (0.235) (0.235) (0.252) (0.252)

PASTGIVING 0.091∗ 0.072(0.045) (0.045)

PASTGIVSQ 0.027∗∗ 0.026∗∗

(0.007) (0.007)

NORDIC 2.560∗∗ 2.846∗∗

(0.421) (0.431)

SOVIET 3.821∗∗ 3.265∗∗

(0.898) (1.111)

YUGOSLAV 4.354∗∗ 3.866∗∗

(0.824) (0.929)

INTERCEPT 0.382 1.634∗∗ -0.532 -0.524 -1.239∗∗

(-0.402) (0.353) (0.412) (0.437) (0.444)

Obs. 12,151 12,151 12,151 10,849 10,849ln L -23,192 -23,211 -23,151 -20,733 -20,696χ2 525.62 475.67 596.75 458.35 522.76ρ 0.145 0.149 0.129 0.118 0.106(s.e.) (0.008) (0.008) (0.007) (0.007) (0.007)

All specifications are estimated using a random-effects tobit model. Standard errors arein parentheses. Significance levels: † : 10%, ∗ : 5%, ∗∗ : 1%.

18

creases the magnitude and significance of the coefficients on the affinity variables. In particular,

the coefficient on PROXIMITY more than doubles in size. Clearly the trade variables were

picking up some of the impact of geographical factors in determining points awarded.

In the descriptive section of the paper we identified some groups of countries that seem

to exchange a lot of points with each other. In order to formally test our conjectures we

defined three clusters of countries: NORDIC, including Sweden, Finland, Norway, Denmark,

Iceland and Estonia; SOVIET, including countries that used to be part of the Soviet union

(Estonia, Latvia, Lithuania, Moldova, Russia, Ukraine); and YUGOSLAV, including countries

that comprised former Yugoslavia (Bosnia-Herzegovina, Croatia, FYR Macedonia, Slovenia,

Serbia & Montenegro). The three dummy variables (taking the value of 1 if both countries

belong to the cluster) were added as explanatory variables in the random-effects tobit regressions.

Column 3 of Table 4 presents the results with the addition of the geographical clusters. For

example the variable NORDIC indicates that a country in the group is likely to receive on

average an additional two and a half points from another country in the group relative to what

it would receive from other countries, regardless of song quality, whereas the case for SOVIET

and YUGOSLAV are even more dramatic. Adding the cluster indicators removes significance

from the other variables, particularly PROXIMITY as can be seen from comparing columns 1

and 3 of Table 4.

Our affinity variables explain a lot of the variation in point-giving but fail to capture two

aspects that may be important. Because they are all time-invariant, we do not capture any

intertemporal changes in affinity. For example, Greece and Turkey exchanged a total of 20

points between them in 14 contests up to 1999. In that year relations between the two countries

improved dramatically as they helped each other when they were successively hit by earthquakes.

In the five contests that followed they exchanged a total of 54 points. Our affinity variables also

do a poor job capturing asymmetries in affinity. This is because all but one of the variables in

vector x take the same value regardless of the direction of giving. The only exception is the

exports variable which takes different values in each direction: EXPORTSij 6= EXPORTSji.

One way to account for asymmetry in affinity is to use voting data from past contests.

19

We define the variable PASTGIVING to be the over-giving by the giver to the receiver in all

previous contests; that is, the difference between the mean number of points that the receiver

has obtained from the giver in the past and what it had received on average from all other

countries. This variable also captures any other factors that determine affinity in a symmetric

fashion but we did not account for. The last two columns of Table 4 present the results with

PASTGIVING and its square with and without the cluster indicators. PASTGIVING and its

square are very significant and indicate that historical asymmetries in one country’s average

voting behavior, when compared with the historical average of everyone else follow a nonlinear

(convex) pattern whereby they get exacerbated the greater they are. Clearly, past asymmetric

behavior strongly affecting present voting goes well beyond a voting rule that rewards “artistic

quality” and strongly points to the presence of networks with strong cultural ties that withstand

time. Note also that the last column dominates in terms of fit (value of the log-likelihood

function) as well as statistical specification, since the tests of joint statistical significance for

various subsets of variables such as the cluster indicators and the PASTGIVING and its square

have zero probability values.

Juries of experts versus public opinion

Among the main important institutional changes in the ESC is the introduction of televoting

in 1998. Table 5 presents the results of the model which is subjected to a structural break in

1998.16 The model includes all the variables that we had earlier as well as their interactions

with the break dummy that takes the value one for observations after and including 1998. We

present two versions, one that includes PASTGIVING and one that does not. In both cases the

interacting variables are jointly significant signifying the importance of the break with a zero

probability value. All affinity variables are more important in the post-1998 period than before.

This suggests that the public is more likely than juries of experts to take their biases into account

when they cast their vote. The host country and order of appearance effects are also much larger16We also considered 1992 as a possible date for a structural break as this was the year that many new countries

from the SOVIET and YUGOSLAV clusters of countries entered the competition. However, the results suggestthat it is 1998 that clearly presented a significant break.

20

Table 5: Test of structural break in 1998Variables pre-1998 post-1998 pre-1998 post-1998EXPORTS 7.087∗∗ -7.638† 5.687∗ -6.067

(2.599) (3.926) (2.710) (4.079)

TOTTRADE 3.272† 4.761† 3.455† 6.741∗

(1.907) (2.848) (2.001) (3.008)

ANYRELIG 0.217 0.447 0.145 0.257(0.207) (0.308) (0.223) (0.328)

ANYLANG -0.296 2.054∗∗ -0.34 2.246∗∗

(0.224) (0.350) (0.235) (0.366)

PROXIMITY 0.071 0.387∗∗ 0.05 0.191(0.079) (0.116) (0.084) (0.126)

HOST 1.206∗∗ 2.467∗∗ 0.956∗ 2.778∗∗

(0.363) (0.649) (0.382) (0.669)

ENGLISH 4.253∗∗ -3.796∗∗ 3.938∗∗ -3.574∗∗

(0.284) (0.361) (0.299) (0.380)

FRENCH 1.521∗∗ -2.241∗∗ 1.251∗∗ -1.934∗∗

(0.256) (0.494) (0.268) (0.513)

FIRST3 -0.819∗∗ -0.143 -1.010∗∗ 0.201(0.233) (0.406) (0.249) (0.423)

LAST3 0.476∗ 1.115∗∗ -0.090 1.995∗∗

(0.227) (0.394) (0.247) (0.417)

SOLOMALE -0.062 -1.256∗∗ 0.230 -1.576∗∗

(0.257) (0.412) (0.288) (0.445)

SOLOFEMALE 1.522∗∗ -0.891∗ 1.509∗∗ -0.829∗

(0.232) (0.382) (0.260) (0.409)

DUET 1.663∗∗ -0.932† 1.753∗∗ -1.004†

(0.287) (0.509) (0.314) (0.537)

PASTGIVING 0.154∗∗ -0.056(0.050) (0.100)

PASTGIVSQ 0.009 0.016(0.009) (0.016)

INTERCEPT 0.071 1.647∗ -0.140 0.291(0.51) (0.735) (0.552) (0.812)

Obs. 12,151 10,849ln L -23,023 -20,576χ2 857.39 760.24ρ 0.132 0.108(s.e.) (0.007) (0.007)

All specifications are estimated using a random-effects tobit model. Standarderrors are in parentheses. Significance levels: † : 10%, ∗ : 5%, ∗∗ : 1%.

21

Table 6: Unbiased rankingsActual contest Counterfactual controlling for:

ranking and points only affinity andaffinity characteristics

1998 contest1 Israel 172 Israel Israel2 UK 166 Malta Belgium3 Malta 165 UK Netherlands4 Netherlands 150 Netherlands Croatia5 Croatia 131 Belgium Malta6 Belgium 122 Croatia Sweden7 Germany 86 Norway Portugal8 Norway 79 Germany Norway9 Ireland 64 Ireland UK

10 Sweden 53 Sweden Germany

2004 contest1 Ukraine 280 Serbia & Mont. Ukraine2 Serbia & Mont. 263 Ukraine Serbia & Mont.3 Greece 252 Greece Cyprus4 Turkey 195 Turkey Sweden5 Cyprus 170 Cyprus Greece6 Sweden 170 Sweden Turkey7 Albania 106 Spain Spain8 Germany 93 Germany Croatia9 Bosnia 91 Bosnia-Herz. France

10 Spain 87 Poland Germany

in the post-1998 period. On the other hand, the effect of language is reversed after 1998. We

think that this is a spurious result reflecting the fact that some of the recent winners did not use

english even though most other countries did. The effect of female performers seems to only be

important in the pre-1998 era and is reversed after that. More importantly, the PASTGIVING

variables are insignificant in the post-1998 era. This is a direct implication of the fact that all

the other variables explain a lot more during this period.

22

Unbiased rankings

An interesting counterfactual experiment is to use the estimated model to construct alternative

rankings after the various sources of bias have been controlled for. We performed this calculation

for two contests, 1998 and 2004, which were selected because the voting was close.17 Table 6

presents the top ten ranked countries according to our counterfactual which are contrasted with

the actual rankings. The first three columns give the final rank, name of the country and points

received. In the fourth column we present the predicted rankings when we control only for

affinity (the x variables) while in the fifth column we also control for song characteristics (the

w variables).18 For 1998, counterfactual rankings from the model that controls only for affinity

coincide with actual rankings in four out of the top ten countries. For the other six there is a

simple interchange of positions. For example, the second and third positions are reversed in the

predictions from the actual rankings (UK and Malta) and the same is true for the fourth and

fifth and seventh and eighth. The reshuffling is even more dramatic when we also control for

song characteristics. Only two out of ten entries (including the winner) remain in their original

positions. The same is true for 2004, where controlling for affinity changes the position of four

countries. Importantly, it places Yugoslavia instead of the Ukraine as the winner (by a very

narrow margin, it must be noted). Adding song characteristics as control variables reverses

the order of the two countries back to the their actual ranking in the contest. The remaining17The 1998 contest in particular was a source of some minor controversy. The following quote from Wikipedia

succinctly describes the last stages of the voting process:

With just one country left to vote, it was anyone’s guess who was going to prevail, with Israel andMalta locked in battle on the same points total (or so the scoreboard said - in fact, Spain’s vote hadbeen wrongly tallied and Malta was really one point ahead), and the United Kingdom apparentlynine points behind (really nine behind Malta and eight behind Israel). When FYR Macedonia cameto award the decisive points, Israel were the first of the three contenders to be mentioned, receivingeight points. That was enough to knock the UK out of contention for victory, but left plenty of roomfor Israel to be overtaken by their principal rival. Next, the ten points went to the UK, nudging theminto what looked like being an extremely fleeting spell in second place, since most of the audienceassumed the twelve points were destined for Malta. Instead, there were gasps as FYR Macedoniasent the final points of the evening to fellow Balkan nation Croatia, handing Israel their first win inthe contest since “Hallelujah” in 1979.

[http://en.wikipedia.org/wiki/Eurovision Song Contest 1998, accessed Feb 2, 2006.]

18We implement the counterfactual by estimating a model with song fixed effects and the variables we want tocontrol for. The estimated fixed effect is our measure of song quality net of the effects captured by the controlvariables.

23

eight countries in the top ten all change positions. The most notable change is the big drop

experienced by the host country, which is a result of the large coefficient on the HOST variable.

This may be excessive punishment because, as we noted in the discussion above, the host country

may receive more points because it makes it a greater effort to make a good showing.

6 Concluding remarks

In this paper we have examined the nature of voting that takes place in the Eurovision song

contest, a festival that offers an example of a truly international forum where a country can

express an opinion about another country, free of political or economic considerations. We find

that the awarding of points to songs goes beyond rewarding a “good” song. As it happens

large differences in voting patterns accounting for tastes reflect some deeper sociological likes

and dislikes among nations that manifest themselves in systematically biased voting. These

systematic biases conceal the different considerations beyond the aesthetic quality of the song

itself that enter the voting preferences of participant countries and are captured by what we

have identified as affinity factors, that is variables that measure how each country feels towards

another country. An interesting observation is that the general public exhibits these biases in

their voting pattern in stronger terms than juries of experts. Of course, one could offer a more

benign interpretation for these apparent biases as reflecting similarities (differences) in tastes,

not necessarily deeper sentiments of likes or dislikes. In any case, we find that affinity variables

are quite important in the voting process in terms of both rankings and points awarded. Using

different specifications we are able to test for the presence and importance of other attributes

of the song delivery that go beyond the affinity variables or objective quality of the song, such

as the order of appearance and the gender of the artist presenting the song and hosting the

competition. These variables are also quite important in explaining voting patterns beyond

pure quality considerations. Our counterfactual rankings that control for affinity and for song

characteristics indicate that these biases add up to a substantial number of points and have an

important effect in the contest’s final outcome.

24

References

Allison, P., and N. Christakis, 1994, “Logit Models for Sets of Ranked Items,” in Peter V.Marsden (ed.), Sociological Methodology, Volume 24 . pp. 199–228, Blackwell, Oxford.

Anderson, J. E., 1979, “A Theoretical Foundation for the Gravity Equation,” American Eco-nomic Review, 69, 106–16.

Beggs, S., S. Cardell, and J. A. Hausman, 1981, “Assessing the Potential Demand for ElectricCars,” Journal of Econometrics, 17, 1–19.

Bernard, A. B., and M. R. Busse, 2004, “Who Wins the Olympic Games: Economic Resourcesand Medal Totals,” Review of Economics and Statistics, 86, 413–417.

Bornhorst, F., A. Ichino, S. Karl, and E. Winter, 2004, “Trust and Trustworthiness amongEuropeans: South-North Comparison,” mimeo.

Doosje, B., and H. S. Alexander, 2005, “What Have they Done for Us Lately? The Dynamicsof Reciprocity in Intergroup Contexts,” Journal of Applied Social Psychology, 35, 508–535.

Fenn, D., O. Suleman, J. Efstathiou, and N. F. Johnson, 2005, “How Does Europe Make ItsMind Up? Connections, Cliques and Compatibility between Countries in the EurovisionSong Contest,” mimeo.

Fershtman, C., and U. Gneezy, 2001, “Discrimination in a Segmented Society: An ExperimentalApproach,” Quarterly Journal of Economics, 116, 351–377.

Gandal, N., C. King, and M. Van Alstyne, 2005, “Information Technology Use and Productivityat the Individual Level,” mimeo.

Ginsburgh, V., and A. Noury, 2004, “Cultural Voting: The Eurovision Song Contest,” mimeo.

Guiso, L., P. Sapienza, and L. Zingales, 2003, “Cultural Biases in Economic Exchange,” mimeo.

Haan, M. A., S. G. Dijkstra, and P. T. Dijkstra, 2005, “Expert Judgement Vesus Public Opinion:Evidence from the Eurovision Song Contest,” Journal of Cultural Economics, 29, 59–78.

Hausman, J. A., and P. A. Ruud, 1987, “Specifying and Testing Econometric Models for Rank-Ordered Data,” Journal of Econometrics, 34, 83–104.

Leamer, E. E., and J. Levinsohn, 1995, “International Trade Theory: The Evidence,” in Gene M.Grossman, and Kenneth Rogoff (ed.), Handbook of International Economics, vol. 3, . pp.1339–1394, North-Holland, Amsterdam.

Marden, J. I., 1995, Analyzing and Modeling Rank Data, Chapman and Hall Publishers, London.

Newman, M., 2003, “The Structure and Function of Complex Networks,” SIAM Review, 45,167–256.

25